Can Claude or ChatGPT AI Do Your Internal Linking? What Works and What Breaks

Partly. Claude and ChatGPT handle the language side of internal linking well: anchor ideas, a reworded sentence, a link plan for a site small enough to paste in. They handle the data side badly: knowing which URLs exist, repeating the same plan twice, seeing a 5,000-page site at once, remembering existing links and publishing changes.

When Ahrefs measured how often AI assistants send people to broken pages, one of the most-visited made-up URLs on its own site was /blog/internal-links/: an internal linking guide at an address that returns a 404. Facts below come from vendor documentation and published studies, checked November 2026.

What can ChatGPT and Claude do well for internal linking?

Both do the writing parts well when you supply every fact in the prompt: give the model the text and the real URLs, and the remaining work is anchor choice and phrasing.

A prompt that stays inside those strengths:

"Below is a paragraph from my article on cold brew ratios, then the URL and title of my grinder guide. Find up to three phrases already in the paragraph that could link to the grinder guide. If none fit, rewrite one sentence. Use only the URL I gave you."

Where does AI internal linking break?

It breaks wherever the job needs site data the chatbot cannot see, keep or act on. Five failure points show up, and each gets worse as the site grows.

1. It invents URLs

Chatbots generate plausible text, and a plausible URL is not always a real one. Ryan Law, Director of Content Marketing at Ahrefs, describes the pattern: "Another portion of cited URLs are true hallucinations, in the sense that they fit the expected pattern of URLs for a given website, but don't actually exist."

In Ahrefs' study of 16 million URLs (https://ahrefs.com/blog/how-often-do-ai-assistants-hallucinate-links) (September 2025), AI assistants sent visitors to 404 pages at 2.87 times the rate of Google Search: 0.43% of clicks versus 0.15%. ChatGPT had the highest rate measured, 1.01% of clicked URLs; Claude came in at 0.58%.

Ahrefs notes the true rate is probably understated, since some 404s are removed pages rather than inventions. Either way, an invented slug is a broken link you publish; ask for "a link to my espresso guide" without the URL, and the model has to guess one.

2. The same prompt gives a different plan each time

Language models sample their output, so repeated runs differ. SparkToro co-founder Rand Fishkin puts it bluntly: "These tools are probability engines: they're designed to generate unique answers every time. Thinking of them as sources of truth or consistency is provably nonsensical."

SparkToro's consistency research (January 2026) ran 12 prompts 2,961 times across ChatGPT, Claude and Google's AI, and found less than a 1 in 100 chance that ChatGPT or Google's AI would return the same list of brands across 100 runs. The study measured brand recommendations, not links, but two runs still give two lists, and neither is a plan you can audit.

3. Large sites don't fit in the chat

A context window is the amount of text a model can consider at once. Current Claude models offer 1M tokens; OpenAI's pricing page lists 54K of total context for GPT Instant on Plus and 27K on Free.

Rough math, an estimate: a 1,500-word article is 2,000 to 2,700 tokens, so 10,000 pages of full text runs 20M to 27M tokens, 20 to 27 times a 1M window. A slim inventory of URL, title and a 150-word summary still runs 2.7M to 3.6M tokens, and titles alone leave the model nothing to judge relevance with.

Bigger windows bring their own problem: Anthropic's context window documentation (https://platform.claude.com/docs/en/build-with-claude/context-windows) warns that "as token count grows, accuracy and recall degrade," an effect it calls context rot.

4. It has no record of your existing links

A chat starts with no map of which anchors point where, which pages are orphans and which links repeat; memory features store notes about you, not a current link graph. Without that record the model can't spot anchor cannibalization, one phrase linking to two different pages, or a source page already linking to the same target three times.

5. It can't publish, or it publishes blind

A plain chat returns text, which is fine for 10 links pasted in by hand and miserable for 300. General-purpose CMS connectors close part of that gap, but a connector that can edit posts doesn't know your link graph: it can insert a link whose anchor already points somewhere else, which makes the same mistakes faster.

The fix: connect the AI to a real link database

The fix is to let the assistant query a tool that already holds your site's pages, links and search data. The Model Context Protocol (MCP), an open standard Anthropic introduced in November 2024, is how most AI apps now connect to tools like that: the model asks, the tool answers with real URLs and real sentences, and the model's job shrinks to choosing, explaining and getting your approval.

Failure Plain chat AI connected to a linking database
Invented URLs Guesses slugs it hasn't seen Picks from URLs the tool returns
Run-to-run drift New list every time One index, changes logged
Context limits Sees only what you paste Queries the index page by page
No link record Starts every session blank Reads current links and anchors
Publishing Manual copy and paste Writes after a preview you approve

MCP doesn't remove every limit: tool results still count against the context window, and Claude Code caps MCP output at 25,000 tokens by default. That is why a well-built server returns short, targeted answers instead of whole pages.

Worked example: LinkBoss MCP

LinkBoss is an AI-powered internal linking tool that builds semantically relevant links in bulk across WordPress, Shopify and other CMSs, used by over 8,000 SEOs. Its internal linking MCP server (https://linkboss.io/internal-linking-mcp-server/) connects Claude (web, desktop and Claude Code), ChatGPT, Cursor, VS Code with GitHub Copilot and other remote-MCP apps to that link data, on any website.

Through it, the assistant reads every page's links and anchors, sees Search Console data (pages close to page one, declining pages), finds opportunities with the exact sentence and an anchor score, builds silos, and fixes anchor cannibalization, http:// links and repeated links.

Its guardrails target the failure points above: every change starts as a dry run showing the linked words in [brackets] plus the credit cost, it adds at most 25 links per call, refuses an anchor that already links to a different page, and logs every change in Link History. Changes go live in about a minute.

MCP access needs a paid subscription, not a free plan, trial or credit pack; plans start at $11 a month for Launch (200 credits, one site). Reading and fixing existing links is free; each added link costs 1 or 2 credits. ChatGPT gets full access only on Business, Enterprise and Edu (Pro is read-only). Any website works: WordPress and Shopify insert links directly, and on other platforms the assistant you already use to update the site applies the approved links through the same server.

Other routes: linking MCPs and auto-linking plugins

LinkBoss isn't the only route. Silotree's linking MCP runs read-only or read-write, previews changes and works on every plan, including free; SEO Cartograph's MCP files AI proposals in a dashboard you apply yourself. Without a chatbot, Link Whisper suggests WordPress links you approve one by one, and LinkStorm suggests links on any CMS but injects them with JavaScript, so they vanish if you cancel.

Pick by write access: proposal-only tools are most cautious; write-capable tools save the most time, provided every write passes a preview you approve.

Which approach fits your site size?

The bigger the site, the less a plain chat can do and the more the AI needs a link database. Treat the bands below as rules of thumb.

Site size Best approach Who places the links Main risk to watch
Under 50 pages Plain chat with your URL list pasted in You, by hand Typos and invented URLs
50 to 500 pages Chat for anchors and rewrites, plus an audit tool or auto-linking plugin You, or a linking tool with review Orphan pages the chat can't see
500 to 5,000 pages AI connected to a linking MCP, in reviewed batches The tool, after your approval Approving dry runs without reading them
5,000+ pages or several client sites Linking MCP or bulk linking tool, with Search Console priorities The tool, in logged batches Over-linking and anchor conflicts at scale

Two questions settle most cases: can the whole site fit in one chat with enough content to judge relevance, and how many links will you place each month? If the answers are "no" and "more than I'd paste by hand," connect the AI to a linking tool.

FAQ

Can ChatGPT add internal links to my WordPress site by itself? Not from a plain chat, which only returns text. Through MCP it can, and OpenAI says ChatGPT requires manual confirmation before write actions; full custom MCP access needs a Business, Enterprise or Edu plan.

Is Claude better than ChatGPT for internal linking? In Ahrefs' data, Claude sent users to 404 pages less often (0.58% versus 1.01%), and SparkToro found Claude slightly more consistent. Both gaps are small next to the gap between guessing and reading real link data, so use whichever app your team already works in.

Is there an AI auto-linking plugin for WordPress that works with ChatGPT or Claude? Through MCP, yes: LinkBoss and Silotree both offer linking servers Claude and ChatGPT can connect to. Plugins such as Link Whisper run their own suggestions and auto-linking without a chatbot.

Do I need a paid plan to use AI for internal linking through MCP? It depends on both sides: LinkBoss's MCP comes with paid subscriptions only, while Silotree says its MCP works on every plan, including free. On the app side, Claude's free plan allows one custom connector and ChatGPT needs Business, Enterprise or Edu for full access.